Debate aids weak judges only if critics outperform them
research on ai debate found weak judges only identify correct answers better if the criticizing model surpasses the judge's own ability to classify errors.
a study pairing models like claude opus 4.6 with weaker judges like opus 4.5 found debate outperforms consultancy by roughly 14-16 percentage points in f1 score, but only under specific conditions. debate succeeded when the critic model exceeded the judge's classification capabilities and the judge actively verified the critique rather than treating it as authority. researchers found the benefit came primarily from the first critique message, with rebuttal rounds adding little value during inference. results suggest debate is a partial solution for scalable oversight, requiring audits to ensure critics can actually identify errors that judges would miss.